Changes touching this path

• the REPRESENTATION is extracted and the PREDICATE is not, because the overlap between a pathspec and a view is exactly the storage while the difference is exactly the RULE: a pathspec bare docs also keeps docs/adr/x.md, the git dash-dash affordance, and a view line has only the glob. that difference must STAY, and not because the two live in different places - a view line is STANDING MATERIALIZATION POLICY, so an implicit slash-star-star would silently WIDEN every one-word line already sitting in somebody .loot/view, and the widening would surface as FILES APPEARING ON DISK rather than as an error. the affordance is right where an operator types a spec and reads the answer in one breath and wrong for stored policy. the decisive evidence for the shape is that policy.rs ALREADY SOLVES THIS TWICE - IgnoreLine and Rule each pair a glob with its text as written in one struct - so Pathspec and View were the two OUTLIERS rather than a new abstraction being invented. desynchronisation is now impossible BY THE COMPILER rather than by a comment: the pair fields are private to policy, the list is private, and push is the only place a Pattern is built, taking ONE string and deriving both halves - proven by two mutations that FAIL TO COMPILE, E0616 reaching for the private list and E0451 constructing a mismatched pair. the prose invariant that said the two cannot disagree because new pushes them together - exactly the comment this ticket exists to remove - now points at the structure. one allocation per pattern per path leaves the match loop, a format in a loop that runs once per tree entry becomes a strip_prefix, and NO SPEEDUP IS CLAIMED ANYWHERE because nothing was measured: the machine was building throughout, which is the CPU condition that makes the harness load arm skip SILENTLY. six mutations with counts read, and the oracle was checked rather than assumed - which found TWO UNPINNED BEHAVIOURS the pre-existing suite was green under: that the prefix arm stops at a whole path COMPONENT, where every other negative case in the module missed by more than one character, and the rule difference itself, now one test that reddens if the two are unified in EITHER direction (#1552) fe21c932 · dbf3dbe6…

Renames are not followed. loot's tree maps a path to an address, so a rename is a delete and an add. This list is the history of the name, not of the bytes.